Evolutionary Planner/Navigator: Operator Performance and Self-Tuning

Based on evolutionary computation concepts, the Evolutionary Planner/Navigator (EP/N) 4, 5] represents a new approach to path planning and navigation. Since its rst version, the development of the EP/N system has been an ever living \evolution" process itself: much new development and further research has been made 8, 7] to fullll the EP/N promise of being able to (1) accommodate diierent optimization criteria , (2) achieve both near-optimality of paths and high planning eeciency, (3) be exible to changes, and (4) be robust to uncertainties. A more important promise of the EP/N is its ability for performance self-tuning to adapt to diierent task environments, mostly through the adaptiveness of its genetic operations. This paper introduces a methodology to measure the overall performance of the EP/N operators and demonstrates how such a measure, called`performance index' for each operator, can be used to make the EP/N adaptive. 1 Background The motion planning problem for mobile robots is typically formulated as follows 9]: given a robot and a description of an environment, plan a path of the robot between two speciied locations, which is collision-free and satisses certain optimization criteria. Although a great deal of research has been done in motion planning and navigation (see 9, 2] for surveys), conventional approaches tend to be innexible in responding to (1) diierent optimization goals and changes of goals, (2) diierent environments or changes and uncertainties in an environment, and (3) diierent constraints on computational resources (such as time and space). Traditional planners either try to search for the optimal path based on some xed criteria (most commonly the shortest path) by any costs or simply do not try to optimize a path. There is a need to have a general, exible, and even adaptive planner capable of meeting any changes in requirements and environments. The EP/N system 4, 5, 7, 8] has been developed to meet such a need, inspired by the following ideas/observations: (1) randomized search can be the most eeective in dealing with NP-hard problems and in escaping local minima; (2) parallel search actions not only increase speed but also provide ground for interactions among search actions to achieve even greater eeciency in optimization; (3) creative application of the evolutionary computation concept to incorporate heuristic knowledge is more eeective in solving practical problems than dogmatic imposition of a standard algorithm; (4) intelligent behavior is the result of a collection of simple reactions to …